Please consider vosk-rs for CPU-only real-time voice recognition

[gtubolcev]

Hi, I hope it's a right place to ask for new voice recognition engines.
vosk-rs is a Rust binding for vosk engine (the latter seems to be a part of Kaldi project).
It supports CPU-only real-time recognition with an accuracy comparable to whisper and others.
Just look at this example here: https://fars.ee/F9-b.mp4

https://github.com/wzhd/vosk-rs
https://github.com/alphacep/vosk-tts
https://github.com/nshmyrev seems to be the developer behind vosk.

Thank you for your work )

[gtubolcev]

FYI there is a leaderboard for russian language ASR with different engines compared. Here is a google translation
https://alphacephei-com.translate.goog/nsh/2025/04/18/russian-models.html?_x_tr_sl=ru&_x_tr_tl=en&_x_tr_hl=de&_x_tr_pto=wapp#
NB Google shows "wax 0.54" instead of "vosk 0.54"

[cjpais]

@gtubolcev curious have you used any of these models yourself? Mostly just to get an idea of performance before implementing anything. It does look potentially quite straightforward but I have limited time. Definitely will accept PRs as well

[gtubolcev]

@cjpais Hi,
I have used some of these models. Vosk gives you a good recognition rate running on a CPU only. That's why there are dozens of projects on github using vosk as the main voice recognition engine for real-time recognition using very small resources. I am not sure if vosk is the better way or the next-gen Kaldi project (sherpa-onnx) etc.
I tried another one from this list, GIGA-AM, for video subtitles generation. It is very good for russian. I don't know if it works as good for other languages.
Lets ask mr. shmyrev.